Novel Staging fixes who is where at the start of each scene. Each video segment of an AI short drama is generated on its own, so group scenes drift: guests appear at a table, and the wrong person ends up seated. This skill plans the opening arrangement once, before the storyboard.
It reads scene images to identify things people can stand beside or sit on, projects them onto a top-view plan, assigns each character a place, posture and facing, and adds extras only when the script calls for them. The default result is an HTML report where you drag people on the plan and export the edited JSON. Eight quality gates run as code. The skill is written in Chinese and reports default to Chinese, with an English option.
What You Get
- A staging.json with items per location and opening positions per scene
- An interactive HTML report with a draggable top view and live checks
- One blocking sentence per scene, ready to paste into the first storyboard segment
- An optional Markdown export when a text version is wanted
Common Use Cases
- Seating six characters around a table so every generated segment agrees
- Marking who is absent when a scene opens and who walks in later
- Adjusting a position with one sentence such as moving a character to the bow of a boat
What Makes It Unique
Positions may only land on items that were identified in the scene image, so nobody sits on a platform that exists only in a prompt. Directions are given as compass sides of objects, not left and right of the screen, because a camera change would flip the latter. The blocking sentence is generated by code and copied word for word into the storyboard.
Under The Hood
The skill is an instruction file with Node.js scripts that use only the standard library. It runs shell commands to seed, validate, render and print blocking sentences, and it starts sub-agents to read images and plan scenes in parallel when the environment allows. It reads the script, art and outline JSON files and the scene images in an images folder, and writes JSON and HTML files. The generated report runs a script in your browser for dragging and exporting, and it stores edits in the browser. It does not fetch web pages or call outside APIs and needs no API key.